AI/text-generation-webui
1
0
Fork 0
mirror of https://github.com/oobabooga/text-generation-webui.git synced 2024-10-01 01:26:03 -04:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #6118 from oobabooga/dev

Browse Source 
Merge dev branch
This commit is contained in:
oobabooga 2024-06-13 00:38:03 -03:00 committed by GitHub
commit 4820ae9aef
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
37 changed files with 481 additions and 695 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

272
README.md
View File

@ -11,7 +11,7 @@ Its goal is to become the [AUTOMATIC1111/stable-diffusion-webui](https://github.
## Features
* 3 interface modes: default (two columns), notebook, and chat.
* Multiple model backends: [Transformers](https://github.com/huggingface/transformers), [llama.cpp](https://github.com/ggerganov/llama.cpp) (through [llama-cpp-python](https://github.com/abetlen/llama-cpp-python)), [ExLlamaV2](https://github.com/turboderp/exllamav2), [AutoGPTQ](https://github.com/PanQiWei/AutoGPTQ), [AutoAWQ](https://github.com/casper-hansen/AutoAWQ), [GPTQ-for-LLaMa](https://github.com/qwopqwop200/GPTQ-for-LLaMa), [QuIP#](https://github.com/Cornell-RelaxML/quip-sharp).
* Multiple model backends: [Transformers](https://github.com/huggingface/transformers), [llama.cpp](https://github.com/ggerganov/llama.cpp) (through [llama-cpp-python](https://github.com/abetlen/llama-cpp-python)), [ExLlamaV2](https://github.com/turboderp/exllamav2), [AutoGPTQ](https://github.com/PanQiWei/AutoGPTQ), [AutoAWQ](https://github.com/casper-hansen/AutoAWQ).
* Dropdown menu for quickly switching between different models.
* Large number of extensions (built-in and user-contributed), including Coqui TTS for realistic voice outputs, Whisper STT for voice inputs, translation, [multimodal pipelines](https://github.com/oobabooga/text-generation-webui/tree/main/extensions/multimodal), vector databases, Stable Diffusion integration, and a lot more. See [the wiki](https://github.com/oobabooga/text-generation-webui/wiki/07-%E2%80%90-Extensions) and [the extensions directory](https://github.com/oobabooga/text-generation-webui-extensions) for details.
* [Chat with custom characters](https://github.com/oobabooga/text-generation-webui/wiki/03-%E2%80%90-Parameters-Tab#character).
@ -200,168 +200,146 @@ pip install -r <requirements file that you have used> --upgrade
List of command-line flags
</summary>
#### Basic settings
```txt
usage: server.py [-h] [--multi-user] [--character CHARACTER] [--model MODEL] [--lora LORA [LORA ...]] [--model-dir MODEL_DIR] [--lora-dir LORA_DIR] [--model-menu] [--settings SETTINGS]
[--extensions EXTENSIONS [EXTENSIONS ...]] [--verbose] [--chat-buttons] [--idle-timeout IDLE_TIMEOUT] [--loader LOADER] [--cpu] [--auto-devices]
[--gpu-memory GPU_MEMORY [GPU_MEMORY ...]] [--cpu-memory CPU_MEMORY] [--disk] [--disk-cache-dir DISK_CACHE_DIR] [--load-in-8bit] [--bf16] [--no-cache] [--trust-remote-code]
[--force-safetensors] [--no_use_fast] [--use_flash_attention_2] [--load-in-4bit] [--use_double_quant] [--compute_dtype COMPUTE_DTYPE] [--quant_type QUANT_TYPE] [--flash-attn]
[--tensorcores] [--n_ctx N_CTX] [--threads THREADS] [--threads-batch THREADS_BATCH] [--no_mul_mat_q] [--n_batch N_BATCH] [--no-mmap] [--mlock] [--n-gpu-layers N_GPU_LAYERS]
[--tensor_split TENSOR_SPLIT] [--numa] [--logits_all] [--no_offload_kqv] [--cache-capacity CACHE_CAPACITY] [--row_split] [--streaming-llm] [--attention-sink-size ATTENTION_SINK_SIZE]
[--gpu-split GPU_SPLIT] [--autosplit] [--max_seq_len MAX_SEQ_LEN] [--cfg-cache] [--no_flash_attn] [--cache_8bit] [--cache_4bit] [--num_experts_per_token NUM_EXPERTS_PER_TOKEN]
[--triton] [--no_inject_fused_mlp] [--no_use_cuda_fp16] [--desc_act] [--disable_exllama] [--disable_exllamav2] [--wbits WBITS] [--groupsize GROUPSIZE] [--no_inject_fused_attention]
[--hqq-backend HQQ_BACKEND] [--deepspeed] [--nvme-offload-dir NVME_OFFLOAD_DIR] [--local_rank LOCAL_RANK] [--alpha_value ALPHA_VALUE] [--rope_freq_base ROPE_FREQ_BASE]
[--compress_pos_emb COMPRESS_POS_EMB] [--listen] [--listen-port LISTEN_PORT] [--listen-host LISTEN_HOST] [--share] [--auto-launch] [--gradio-auth GRADIO_AUTH]
[--gradio-auth-path GRADIO_AUTH_PATH] [--ssl-keyfile SSL_KEYFILE] [--ssl-certfile SSL_CERTFILE] [--api] [--public-api] [--public-api-id PUBLIC_API_ID] [--api-port API_PORT]
[--api-key API_KEY] [--admin-key ADMIN_KEY] [--nowebui] [--multimodal-pipeline MULTIMODAL_PIPELINE] [--model_type MODEL_TYPE] [--pre_layer PRE_LAYER [PRE_LAYER ...]]
[--checkpoint CHECKPOINT] [--monkey-patch]
| Flag | Description |
|--------------------------------------------|-------------|
| `-h`, `--help` | show this help message and exit |
| `--multi-user` | Multi-user mode. Chat histories are not saved or automatically loaded. WARNING: this is likely not safe for sharing publicly. |
| `--character CHARACTER` | The name of the character to load in chat mode by default. |
| `--model MODEL` | Name of the model to load by default. |
| `--lora LORA [LORA ...]` | The list of LoRAs to load. If you want to load more than one LoRA, write the names separated by spaces. |
| `--model-dir MODEL_DIR` | Path to directory with all the models. |
| `--lora-dir LORA_DIR` | Path to directory with all the loras. |
| `--model-menu` | Show a model menu in the terminal when the web UI is first launched. |
| `--settings SETTINGS_FILE` | Load the default interface settings from this yaml file. See `settings-template.yaml` for an example. If you create a file called `settings.yaml`, this file will be loaded by default without the need to use the `--settings` flag. |
| `--extensions EXTENSIONS [EXTENSIONS ...]` | The list of extensions to load. If you want to load more than one extension, write the names separated by spaces. |
| `--verbose` | Print the prompts to the terminal. |
| `--chat-buttons` | Show buttons on the chat tab instead of a hover menu. |
Text generation web UI
#### Model loader
options:
-h, --help show this help message and exit
| Flag | Description |
|--------------------------------------------|-------------|
| `--loader LOADER` | Choose the model loader manually, otherwise, it will get autodetected. Valid options: Transformers, llama.cpp, llamacpp_HF, ExLlamav2_HF, ExLlamav2, AutoGPTQ, AutoAWQ, GPTQ-for-LLaMa, QuIP#. |
Basic settings:
--multi-user Multi-user mode. Chat histories are not saved or automatically loaded. Warning: this is likely not safe for sharing publicly.
--character CHARACTER The name of the character to load in chat mode by default.
--model MODEL Name of the model to load by default.
--lora LORA [LORA ...] The list of LoRAs to load. If you want to load more than one LoRA, write the names separated by spaces.
--model-dir MODEL_DIR Path to directory with all the models.
--lora-dir LORA_DIR Path to directory with all the loras.
--model-menu Show a model menu in the terminal when the web UI is first launched.
--settings SETTINGS Load the default interface settings from this yaml file. See settings-template.yaml for an example. If you create a file called settings.yaml, this
file will be loaded by default without the need to use the --settings flag.
--extensions EXTENSIONS [EXTENSIONS ...] The list of extensions to load. If you want to load more than one extension, write the names separated by spaces.
--verbose Print the prompts to the terminal.
--chat-buttons Show buttons on the chat tab instead of a hover menu.
--idle-timeout IDLE_TIMEOUT Unload model after this many minutes of inactivity. It will be automatically reloaded when you try to use it again.
#### Accelerate/transformers
Model loader:
--loader LOADER Choose the model loader manually, otherwise, it will get autodetected. Valid options: Transformers, llama.cpp, llamacpp_HF, ExLlamav2_HF, ExLlamav2,
AutoGPTQ, AutoAWQ.
| Flag | Description |
|---------------------------------------------|-------------|
| `--cpu` | Use the CPU to generate text. Warning: Training on CPU is extremely slow. |
| `--auto-devices` | Automatically split the model across the available GPU(s) and CPU. |
| `--gpu-memory GPU_MEMORY [GPU_MEMORY ...]` | Maximum GPU memory in GiB to be allocated per GPU. Example: --gpu-memory 10 for a single GPU, --gpu-memory 10 5 for two GPUs. You can also set values in MiB like --gpu-memory 3500MiB. |
| `--cpu-memory CPU_MEMORY` | Maximum CPU memory in GiB to allocate for offloaded weights. Same as above. |
| `--disk` | If the model is too large for your GPU(s) and CPU combined, send the remaining layers to the disk. |
| `--disk-cache-dir DISK_CACHE_DIR` | Directory to save the disk cache to. Defaults to "cache". |
| `--load-in-8bit` | Load the model with 8-bit precision (using bitsandbytes). |
| `--bf16` | Load the model with bfloat16 precision. Requires NVIDIA Ampere GPU. |
| `--no-cache` | Set `use_cache` to `False` while generating text. This reduces VRAM usage slightly, but it comes at a performance cost. |
| `--trust-remote-code` | Set `trust_remote_code=True` while loading the model. Necessary for some models. |
| `--no_use_fast` | Set use_fast=False while loading the tokenizer (it's True by default). Use this if you have any problems related to use_fast. |
| `--use_flash_attention_2` | Set use_flash_attention_2=True while loading the model. |
Transformers/Accelerate:
--cpu Use the CPU to generate text. Warning: Training on CPU is extremely slow.
--auto-devices Automatically split the model across the available GPU(s) and CPU.
--gpu-memory GPU_MEMORY [GPU_MEMORY ...] Maximum GPU memory in GiB to be allocated per GPU. Example: --gpu-memory 10 for a single GPU, --gpu-memory 10 5 for two GPUs. You can also set values
in MiB like --gpu-memory 3500MiB.
--cpu-memory CPU_MEMORY Maximum CPU memory in GiB to allocate for offloaded weights. Same as above.
--disk If the model is too large for your GPU(s) and CPU combined, send the remaining layers to the disk.
--disk-cache-dir DISK_CACHE_DIR Directory to save the disk cache to. Defaults to "cache".
--load-in-8bit Load the model with 8-bit precision (using bitsandbytes).
--bf16 Load the model with bfloat16 precision. Requires NVIDIA Ampere GPU.
--no-cache Set use_cache to False while generating text. This reduces VRAM usage slightly, but it comes at a performance cost.
--trust-remote-code Set trust_remote_code=True while loading the model. Necessary for some models.
--force-safetensors Set use_safetensors=True while loading the model. This prevents arbitrary code execution.
--no_use_fast Set use_fast=False while loading the tokenizer (it's True by default). Use this if you have any problems related to use_fast.
--use_flash_attention_2 Set use_flash_attention_2=True while loading the model.
#### bitsandbytes 4-bit
bitsandbytes 4-bit:
--load-in-4bit Load the model with 4-bit precision (using bitsandbytes).
--use_double_quant use_double_quant for 4-bit.
--compute_dtype COMPUTE_DTYPE compute dtype for 4-bit. Valid options: bfloat16, float16, float32.
--quant_type QUANT_TYPE quant_type for 4-bit. Valid options: nf4, fp4.
⚠️ Requires minimum compute of 7.0 on Windows at the moment.
llama.cpp:
--flash-attn Use flash-attention.
--tensorcores Use llama-cpp-python compiled with tensor cores support. This increases performance on RTX cards. NVIDIA only.
--n_ctx N_CTX Size of the prompt context.
--threads THREADS Number of threads to use.
--threads-batch THREADS_BATCH Number of threads to use for batches/prompt processing.
--no_mul_mat_q Disable the mulmat kernels.
--n_batch N_BATCH Maximum number of prompt tokens to batch together when calling llama_eval.
--no-mmap Prevent mmap from being used.
--mlock Force the system to keep the model in RAM.
--n-gpu-layers N_GPU_LAYERS Number of layers to offload to the GPU.
--tensor_split TENSOR_SPLIT Split the model across multiple GPUs. Comma-separated list of proportions. Example: 18,17.
--numa Activate NUMA task allocation for llama.cpp.
--logits_all Needs to be set for perplexity evaluation to work. Otherwise, ignore it, as it makes prompt processing slower.
--no_offload_kqv Do not offload the K, Q, V to the GPU. This saves VRAM but reduces the performance.
--cache-capacity CACHE_CAPACITY Maximum cache capacity (llama-cpp-python). Examples: 2000MiB, 2GiB. When provided without units, bytes will be assumed.
--row_split Split the model by rows across GPUs. This may improve multi-gpu performance.
--streaming-llm Activate StreamingLLM to avoid re-evaluating the entire prompt when old messages are removed.
--attention-sink-size ATTENTION_SINK_SIZE StreamingLLM: number of sink tokens. Only used if the trimmed prompt does not share a prefix with the old prompt.
| Flag | Description |
|---------------------------------------------|-------------|
| `--load-in-4bit` | Load the model with 4-bit precision (using bitsandbytes). |
| `--use_double_quant` | use_double_quant for 4-bit. |
| `--compute_dtype COMPUTE_DTYPE` | compute dtype for 4-bit. Valid options: bfloat16, float16, float32. |
| `--quant_type QUANT_TYPE` | quant_type for 4-bit. Valid options: nf4, fp4. |
ExLlamaV2:
--gpu-split GPU_SPLIT Comma-separated list of VRAM (in GB) to use per GPU device for model layers. Example: 20,7,7.
--autosplit Autosplit the model tensors across the available GPUs. This causes --gpu-split to be ignored.
--max_seq_len MAX_SEQ_LEN Maximum sequence length.
--cfg-cache ExLlamav2_HF: Create an additional cache for CFG negative prompts. Necessary to use CFG with that loader.
--no_flash_attn Force flash-attention to not be used.
--cache_8bit Use 8-bit cache to save VRAM.
--cache_4bit Use Q4 cache to save VRAM.
--num_experts_per_token NUM_EXPERTS_PER_TOKEN Number of experts to use for generation. Applies to MoE models like Mixtral.
#### llama.cpp
AutoGPTQ:
--triton Use triton.
--no_inject_fused_mlp Triton mode only: disable the use of fused MLP, which will use less VRAM at the cost of slower inference.
--no_use_cuda_fp16 This can make models faster on some systems.
--desc_act For models that do not have a quantize_config.json, this parameter is used to define whether to set desc_act or not in BaseQuantizeConfig.
--disable_exllama Disable ExLlama kernel, which can improve inference speed on some systems.
--disable_exllamav2 Disable ExLlamav2 kernel.
--wbits WBITS Load a pre-quantized model with specified precision in bits. 2, 3, 4 and 8 are supported.
--groupsize GROUPSIZE Group size.
| Flag | Description |
|-------------|-------------|
| `--tensorcores` | Use llama-cpp-python compiled with tensor cores support. This increases performance on RTX cards. NVIDIA only. |
| `--flash-attn` | Use flash-attention. |
| `--n_ctx N_CTX` | Size of the prompt context. |
| `--threads` | Number of threads to use. |
| `--threads-batch THREADS_BATCH` | Number of threads to use for batches/prompt processing. |
| `--no_mul_mat_q` | Disable the mulmat kernels. |
| `--n_batch` | Maximum number of prompt tokens to batch together when calling llama_eval. |
| `--no-mmap` | Prevent mmap from being used. |
| `--mlock` | Force the system to keep the model in RAM. |
| `--n-gpu-layers N_GPU_LAYERS` | Number of layers to offload to the GPU. |
| `--tensor_split TENSOR_SPLIT` | Split the model across multiple GPUs. Comma-separated list of proportions. Example: 18,17. |
| `--numa` | Activate NUMA task allocation for llama.cpp. |
| `--logits_all`| Needs to be set for perplexity evaluation to work. Otherwise, ignore it, as it makes prompt processing slower. |
| `--no_offload_kqv` | Do not offload the K, Q, V to the GPU. This saves VRAM but reduces the performance. |
| `--cache-capacity CACHE_CAPACITY` | Maximum cache capacity (llama-cpp-python). Examples: 2000MiB, 2GiB. When provided without units, bytes will be assumed. |
| `--row_split` | Split the model by rows across GPUs. This may improve multi-gpu performance. |
| `--streaming-llm` | Activate StreamingLLM to avoid re-evaluating the entire prompt when old messages are removed. |
| `--attention-sink-size ATTENTION_SINK_SIZE` | StreamingLLM: number of sink tokens. Only used if the trimmed prompt doesn't share a prefix with the old prompt. |
AutoAWQ:
--no_inject_fused_attention Disable the use of fused attention, which will use less VRAM at the cost of slower inference.
#### ExLlamav2
HQQ:
--hqq-backend HQQ_BACKEND Backend for the HQQ loader. Valid options: PYTORCH, PYTORCH_COMPILE, ATEN.
| Flag | Description |
|------------------|-------------|
|`--gpu-split` | Comma-separated list of VRAM (in GB) to use per GPU device for model layers. Example: 20,7,7. |
|`--max_seq_len MAX_SEQ_LEN` | Maximum sequence length. |
|`--cfg-cache` | ExLlamav2_HF: Create an additional cache for CFG negative prompts. Necessary to use CFG with that loader. |
|`--no_flash_attn` | Force flash-attention to not be used. |
|`--cache_8bit` | Use 8-bit cache to save VRAM. |
|`--cache_4bit` | Use Q4 cache to save VRAM. |
|`--num_experts_per_token NUM_EXPERTS_PER_TOKEN` | Number of experts to use for generation. Applies to MoE models like Mixtral. |
DeepSpeed:
--deepspeed Enable the use of DeepSpeed ZeRO-3 for inference via the Transformers integration.
--nvme-offload-dir NVME_OFFLOAD_DIR DeepSpeed: Directory to use for ZeRO-3 NVME offloading.
--local_rank LOCAL_RANK DeepSpeed: Optional argument for distributed setups.
#### AutoGPTQ
RoPE:
--alpha_value ALPHA_VALUE Positional embeddings alpha factor for NTK RoPE scaling. Use either this or compress_pos_emb, not both.
--rope_freq_base ROPE_FREQ_BASE If greater than 0, will be used instead of alpha_value. Those two are related by rope_freq_base = 10000 * alpha_value ^ (64 / 63).
--compress_pos_emb COMPRESS_POS_EMB Positional embeddings compression factor. Should be set to (context length) / (model's original context length). Equal to 1/rope_freq_scale.
| Flag | Description |
|------------------|-------------|
| `--triton` | Use triton. |
| `--no_inject_fused_attention` | Disable the use of fused attention, which will use less VRAM at the cost of slower inference. |
| `--no_inject_fused_mlp` | Triton mode only: disable the use of fused MLP, which will use less VRAM at the cost of slower inference. |
| `--no_use_cuda_fp16` | This can make models faster on some systems. |
| `--desc_act` | For models that don't have a quantize_config.json, this parameter is used to define whether to set desc_act or not in BaseQuantizeConfig. |
| `--disable_exllama` | Disable ExLlama kernel, which can improve inference speed on some systems. |
| `--disable_exllamav2` | Disable ExLlamav2 kernel. |
Gradio:
--listen Make the web UI reachable from your local network.
--listen-port LISTEN_PORT The listening port that the server will use.
--listen-host LISTEN_HOST The hostname that the server will use.
--share Create a public URL. This is useful for running the web UI on Google Colab or similar.
--auto-launch Open the web UI in the default browser upon launch.
--gradio-auth GRADIO_AUTH Set Gradio authentication password in the format "username:password". Multiple credentials can also be supplied with "u1:p1,u2:p2,u3:p3".
--gradio-auth-path GRADIO_AUTH_PATH Set the Gradio authentication file path. The file should contain one or more user:password pairs in the same format as above.
--ssl-keyfile SSL_KEYFILE The path to the SSL certificate key file.
--ssl-certfile SSL_CERTFILE The path to the SSL certificate cert file.
#### GPTQ-for-LLaMa
API:
--api Enable the API extension.
--public-api Create a public URL for the API using Cloudfare.
--public-api-id PUBLIC_API_ID Tunnel ID for named Cloudflare Tunnel. Use together with public-api option.
--api-port API_PORT The listening port for the API.
--api-key API_KEY API authentication key.
--admin-key ADMIN_KEY API authentication key for admin tasks like loading and unloading models. If not set, will be the same as --api-key.
--nowebui Do not launch the Gradio UI. Useful for launching the API in standalone mode.
| Flag | Description |
|---------------------------|-------------|
| `--wbits WBITS` | Load a pre-quantized model with specified precision in bits. 2, 3, 4 and 8 are supported. |
| `--model_type MODEL_TYPE` | Model type of pre-quantized model. Currently LLaMA, OPT, and GPT-J are supported. |
| `--groupsize GROUPSIZE` | Group size. |
| `--pre_layer PRE_LAYER [PRE_LAYER ...]` | The number of layers to allocate to the GPU. Setting this parameter enables CPU offloading for 4-bit models. For multi-gpu, write the numbers separated by spaces, eg `--pre_layer 30 60`. |
| `--checkpoint CHECKPOINT` | The path to the quantized checkpoint file. If not specified, it will be automatically detected. |
| `--monkey-patch` | Apply the monkey patch for using LoRAs with quantized models. |
#### HQQ
| Flag | Description |
|-------------|-------------|
| `--hqq-backend` | Backend for the HQQ loader. Valid options: PYTORCH, PYTORCH_COMPILE, ATEN. |
#### DeepSpeed
| Flag | Description |
|---------------------------------------|-------------|
| `--deepspeed` | Enable the use of DeepSpeed ZeRO-3 for inference via the Transformers integration. |
| `--nvme-offload-dir NVME_OFFLOAD_DIR` | DeepSpeed: Directory to use for ZeRO-3 NVME offloading. |
| `--local_rank LOCAL_RANK` | DeepSpeed: Optional argument for distributed setups. |
#### RoPE (for llama.cpp, ExLlamaV2, and transformers)
| Flag | Description |
|------------------|-------------|
| `--alpha_value ALPHA_VALUE` | Positional embeddings alpha factor for NTK RoPE scaling. Use either this or `compress_pos_emb`, not both. |
| `--rope_freq_base ROPE_FREQ_BASE` | If greater than 0, will be used instead of alpha_value. Those two are related by `rope_freq_base = 10000 * alpha_value ^ (64 / 63)`. |
| `--compress_pos_emb COMPRESS_POS_EMB` | Positional embeddings compression factor. Should be set to `(context length) / (model's original context length)`. Equal to `1/rope_freq_scale`. |
#### Gradio
| Flag | Description |
|---------------------------------------|-------------|
| `--listen` | Make the web UI reachable from your local network. |
| `--listen-port LISTEN_PORT` | The listening port that the server will use. |
| `--listen-host LISTEN_HOST` | The hostname that the server will use. |
| `--share` | Create a public URL. This is useful for running the web UI on Google Colab or similar. |
| `--auto-launch` | Open the web UI in the default browser upon launch. |
| `--gradio-auth USER:PWD` | Set Gradio authentication password in the format "username:password". Multiple credentials can also be supplied with "u1:p1,u2:p2,u3:p3". |
| `--gradio-auth-path GRADIO_AUTH_PATH` | Set the Gradio authentication file path. The file should contain one or more user:password pairs in the same format as above. |
| `--ssl-keyfile SSL_KEYFILE` | The path to the SSL certificate key file. |
| `--ssl-certfile SSL_CERTFILE` | The path to the SSL certificate cert file. |
#### API
| Flag | Description |
|---------------------------------------|-------------|
| `--api` | Enable the API extension. |
| `--public-api` | Create a public URL for the API using Cloudfare. |
| `--public-api-id PUBLIC_API_ID` | Tunnel ID for named Cloudflare Tunnel. Use together with public-api option. |
| `--api-port API_PORT` | The listening port for the API. |
| `--api-key API_KEY` | API authentication key. |
| `--admin-key ADMIN_KEY` | API authentication key for admin tasks like loading and unloading models. If not set, will be the same as --api-key. |
| `--nowebui` | Do not launch the Gradio UI. Useful for launching the API in standalone mode. |
#### Multimodal
| Flag | Description |
|---------------------------------------|-------------|
| `--multimodal-pipeline PIPELINE` | The multimodal pipeline to use. Examples: `llava-7b`, `llava-13b`. |
Multimodal:
--multimodal-pipeline MULTIMODAL_PIPELINE The multimodal pipeline to use. Examples: llava-7b, llava-13b.
```
</details>

8
docs/04 - Model Tab.md
View File

@ -64,14 +64,6 @@ Loads: GPTQ models.
* **no_use_cuda_fp16**: On some systems, the performance can be very bad with this unset. Can usually be ignored.
* **desc_act**: For ancient models without proper metadata, sets the model "act-order" parameter manually. Can usually be ignored.
### GPTQ-for-LLaMa
Loads: GPTQ models.
Ancient loader, the first one to implement 4-bit quantization. It works on older GPUs for which ExLlamaV2 and AutoGPTQ do not work, and it doesn't work with "act-order", so you should use it with simple 4-bit-128g models.
* **pre_layer**: Used for CPU offloading. The higher the number, the more layers will be sent to the GPU. GPTQ-for-LLaMa CPU offloading was faster than the one implemented in AutoGPTQ the last time I checked.
### llama.cpp
Loads: GGUF models. Note: GGML models have been deprecated and do not work anymore.

2
docs/05 - Training Tab.md
View File

@ -124,7 +124,7 @@ When you're running training, the WebUI's console window will log reports that i
"Loss" in the world of AI training theoretically means "how close is the model to perfect", with `0` meaning "absolutely perfect". This is calculated by measuring the difference between the model outputting exactly the text you're training it to output, and what it actually outputs.
In practice, a good LLM should have a very complex variable range of ideas running in its artificial head, so a loss of `0` would indicate that the model has broken and forgotten to how think about anything other than what you trained it.
In practice, a good LLM should have a very complex variable range of ideas running in its artificial head, so a loss of `0` would indicate that the model has broken and forgotten how to think about anything other than what you trained it on.
So, in effect, Loss is a balancing game: you want to get it low enough that it understands your data, but high enough that it isn't forgetting everything else. Generally, if it goes below `1.0`, it's going to start forgetting its prior memories, and you should stop training. In some cases you may prefer to take it as low as `0.5` (if you want it to be very very predictable). Different goals have different needs, so don't be afraid to experiment and see what works best for you.

22
docs/08 - Additional Tips.md
View File

@ -13,28 +13,6 @@ Source: https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/1126
This file will be automatically detected the next time you start the web UI.
## Using LoRAs with GPTQ-for-LLaMa
This requires using a monkey patch that is supported by this web UI: https://github.com/johnsmith0031/alpaca_lora_4bit
To use it:
Install alpaca_lora_4bit using pip
```
git clone https://github.com/johnsmith0031/alpaca_lora_4bit.git
cd alpaca_lora_4bit
git fetch origin winglian-setup_pip
git checkout winglian-setup_pip
pip install .
```
Start the UI with the --monkey-patch flag:
```
python server.py --model llama-7b-4bit-128g --listen --lora tloen_alpaca-lora-7b --monkey-patch
```
## DeepSpeed
`DeepSpeed ZeRO-3` is an alternative offloading strategy for full-precision (16-bit) transformers models.

2
docs/09 - Docker.md
View File

@ -19,7 +19,7 @@ Use these commands to launch the image:
```
cd text-generation-webui
ln -s docker/{nvidia/Dockerfile,docker-compose.yml,.dockerignore} .
ln -s docker/{nvidia/Dockerfile,nvidia/docker-compose.yml,.dockerignore} .
cp docker/.env.example .env
# Edit .env and set TORCH_CUDA_ARCH_LIST based on your GPU model
docker compose up --build

8
docs/What Works.md
View File

@ -2,15 +2,13 @@
| Loader | Loading 1 LoRA | Loading 2 or more LoRAs | Training LoRAs | Multimodal extension | Perplexity evaluation |
|----------------|----------------|-------------------------|----------------|----------------------|-----------------------|
| Transformers | ✅ | ✅\*\*\* | ✅\* | ✅ | ✅ |
| Transformers | ✅ | ✅\*\* | ✅\* | ✅ | ✅ |
| llama.cpp | ❌ | ❌ | ❌ | ❌ | use llamacpp_HF |
| llamacpp_HF | ❌ | ❌ | ❌ | ❌ | ✅ |
| ExLlamav2_HF | ✅ | ✅ | ❌ | ❌ | ✅ |
| ExLlamav2 | ✅ | ✅ | ❌ | ❌ | use ExLlamav2_HF |
| AutoGPTQ | ✅ | ❌ | ❌ | ✅ | ✅ |
| AutoAWQ | ? | ❌ | ? | ? | ✅ |
| GPTQ-for-LLaMa | ✅\*\* | ✅\*\*\* | ✅ | ✅ | ✅ |
| QuIP# | ? | ? | ? | ? | ✅ |
| HQQ | ? | ? | ? | ? | ✅ |
❌ = not implemented
@ -19,6 +17,4 @@
\* Training LoRAs with GPTQ models also works with the Transformers loader. Make sure to check "auto-devices" and "disable_exllama" before loading the model.
\*\* Requires the monkey-patch. The instructions can be found [here](https://github.com/oobabooga/text-generation-webui/wiki/08-%E2%80%90-Additional-Tips#using-loras-with-gptq-for-llama).
\*\*\* Multi-LoRA in PEFT is tricky and the current implementation does not work reliably in all cases.
\*\* Multi-LoRA in PEFT is tricky and the current implementation does not work reliably in all cases.

12
download-model.py
View File

@ -167,8 +167,11 @@ class ModelDownloader:
is_llamacpp = has_gguf and specific_file is not None
return links, sha256, is_lora, is_llamacpp
def get_output_folder(self, model, branch, is_lora, is_llamacpp=False):
base_folder = 'models' if not is_lora else 'loras'
def get_output_folder(self, model, branch, is_lora, is_llamacpp=False, model_dir=None):
if model_dir:
base_folder = model_dir
else:
base_folder = 'models' if not is_lora else 'loras'
# If the model is of type GGUF, save directly in the base_folder
if is_llamacpp:
@ -304,7 +307,8 @@ if __name__ == '__main__':
parser.add_argument('--threads', type=int, default=4, help='Number of files to download simultaneously.')
parser.add_argument('--text-only', action='store_true', help='Only download text files (txt/json).')
parser.add_argument('--specific-file', type=str, default=None, help='Name of the specific file to download (if not provided, downloads all).')
parser.add_argument('--output', type=str, default=None, help='The folder where the model should be saved.')
parser.add_argument('--output', type=str, default=None, help='Save the model files to this folder.')
parser.add_argument('--model-dir', type=str, default=None, help='Save the model files to a subfolder of this folder instead of the default one (text-generation-webui/models).')
parser.add_argument('--clean', action='store_true', help='Does not resume the previous download.')
parser.add_argument('--check', action='store_true', help='Validates the checksums of model files.')
parser.add_argument('--max-retries', type=int, default=5, help='Max retries count when get error in download time.')
@ -333,7 +337,7 @@ if __name__ == '__main__':
if args.output:
output_folder = Path(args.output)
else:
output_folder = downloader.get_output_folder(model, branch, is_lora, is_llamacpp=is_llamacpp)
output_folder = downloader.get_output_folder(model, branch, is_lora, is_llamacpp=is_llamacpp, model_dir=args.model_dir)
if args.check:
# Check previously downloaded files

4
extensions/openai/typing.py
View File

@ -33,6 +33,10 @@ class GenerationOptions(BaseModel):
seed: int = -1
encoder_repetition_penalty: float = 1
no_repeat_ngram_size: int = 0
dry_multiplier: float = 0
dry_base: float = 1.75
dry_allowed_length: int = 2
dry_sequence_breakers: str = '"\\n", ":", "\\"", "*"'
truncation_length: int = 0
max_tokens_second: int = 0
prompt_lookup_num_tokens: int = 0

13
instruction-templates/Llama-v3.yaml Normal file
View File

@ -0,0 +1,13 @@
instruction_template: |-
{%- set ns = namespace(found=false) -%}
{%- for message in messages -%}
{%- if message['role'] == 'system' -%}
{%- set ns.found = true -%}
{%- endif -%}
{%- endfor -%}
{%- for message in messages -%}
{{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n' + message['content'].rstrip() + '<|eot_id|>' -}}
{%- endfor -%}
{%- if add_generation_prompt -%}
{{-'<|start_header_id|>assistant<|end_header_id|>\n\n'-}}
{%- endif -%}

2
modules/AutoGPTQ_loader.py
View File

@ -44,7 +44,7 @@ def load_quantized(model_name):
'model_basename': pt_path.stem,
'device': "xpu:0" if is_xpu_available() else "cuda:0" if not shared.args.cpu else "cpu",
'use_triton': shared.args.triton,
'inject_fused_attention': not shared.args.no_inject_fused_attention,
'inject_fused_attention': False,
'inject_fused_mlp': not shared.args.no_inject_fused_mlp,
'use_safetensors': use_safetensors,
'trust_remote_code': shared.args.trust_remote_code,

171
modules/GPTQ_loader.py
View File

@ -1,171 +0,0 @@
import inspect
import re
from pathlib import Path
import accelerate
import torch
import transformers
from accelerate.utils import is_xpu_available
from gptq_for_llama import llama_inference_offload
from gptq_for_llama.modelutils import find_layers
from gptq_for_llama.quant import make_quant
from transformers import AutoConfig, AutoModelForCausalLM
import modules.shared as shared
from modules.logging_colors import logger
# This function is a replacement for the load_quant function in the
# GPTQ-for_LLaMa repository. It supports more models and branches.
def _load_quant(model, checkpoint, wbits, groupsize=-1, faster_kernel=False, exclude_layers=None, kernel_switch_threshold=128, eval=True):
exclude_layers = exclude_layers or ['lm_head']
def noop(*args, **kwargs):
pass
config = AutoConfig.from_pretrained(model, trust_remote_code=shared.args.trust_remote_code)
torch.nn.init.kaiming_uniform_ = noop
torch.nn.init.uniform_ = noop
torch.nn.init.normal_ = noop
torch.set_default_dtype(torch.half)
transformers.modeling_utils._init_weights = False
torch.set_default_dtype(torch.half)
model = AutoModelForCausalLM.from_config(config, trust_remote_code=shared.args.trust_remote_code)
torch.set_default_dtype(torch.float)
if eval:
model = model.eval()
layers = find_layers(model)
for name in exclude_layers:
if name in layers:
del layers[name]
gptq_args = inspect.getfullargspec(make_quant).args
make_quant_kwargs = {
'module': model,
'names': layers,
'bits': wbits,
}
if 'groupsize' in gptq_args:
make_quant_kwargs['groupsize'] = groupsize
if 'faster' in gptq_args:
make_quant_kwargs['faster'] = faster_kernel
if 'kernel_switch_threshold' in gptq_args:
make_quant_kwargs['kernel_switch_threshold'] = kernel_switch_threshold
make_quant(**make_quant_kwargs)
del layers
if checkpoint.endswith('.safetensors'):
from safetensors.torch import load_file as safe_load
model.load_state_dict(safe_load(checkpoint), strict=False)
else:
model.load_state_dict(torch.load(checkpoint, weights_only=True), strict=False)
model.seqlen = 2048
return model
# Used to locate the .pt/.safetensors quantized file
def find_quantized_model_file(model_name):
if shared.args.checkpoint:
return Path(shared.args.checkpoint)
path_to_model = Path(f'{shared.args.model_dir}/{model_name}')
pt_path = None
priority_name_list = [
Path(f'{shared.args.model_dir}/{model_name}{hyphen}{shared.args.wbits}bit{group}{ext}')
for group in ([f'-{shared.args.groupsize}g', ''] if shared.args.groupsize > 0 else [''])
for ext in ['.safetensors', '.pt']
for hyphen in ['-', f'/{model_name}-', '/']
]
for path in priority_name_list:
if path.exists():
pt_path = path
break
# If the model hasn't been found with a well-behaved name, pick the last .pt
# or the last .safetensors found in its folder as a last resort
if not pt_path:
for ext in ['.pt', '.safetensors']:
found = list(path_to_model.glob(f"*{ext}"))
if len(found) > 0:
if len(found) > 1:
logger.warning(f'More than one {ext} model has been found. The last one will be selected. It could be wrong.')
pt_path = found[-1]
break
return pt_path
# The function that loads the model in modules/models.py
def load_quantized(model_name):
if shared.args.model_type is None:
logger.error("The model could not be loaded because its type could not be inferred from its name.")
logger.error("Please specify the type manually using the --model_type argument.")
return None
# Select the appropriate load_quant function
model_type = shared.args.model_type.lower()
if shared.args.pre_layer and model_type == 'llama':
load_quant = llama_inference_offload.load_quant
elif model_type in ('llama', 'opt', 'gptj'):
if shared.args.pre_layer:
logger.warning("Ignoring --pre_layer because it only works for llama model type.")
load_quant = _load_quant
else:
logger.error("Unknown pre-quantized model type specified. Only 'llama', 'opt' and 'gptj' are supported")
exit()
# Find the quantized model weights file (.pt/.safetensors)
path_to_model = Path(f'{shared.args.model_dir}/{model_name}')
pt_path = find_quantized_model_file(model_name)
if not pt_path:
logger.error("Could not find the quantized model in .pt or .safetensors format. Exiting.")
exit()
else:
logger.info(f"Found the following quantized model: {pt_path}")
# qwopqwop200's offload
if model_type == 'llama' and shared.args.pre_layer:
if len(shared.args.pre_layer) == 1:
pre_layer = shared.args.pre_layer[0]
else:
pre_layer = shared.args.pre_layer
model = load_quant(str(path_to_model), str(pt_path), shared.args.wbits, shared.args.groupsize, pre_layer)
else:
threshold = False if model_type == 'gptj' else 128
model = load_quant(str(path_to_model), str(pt_path), shared.args.wbits, shared.args.groupsize, kernel_switch_threshold=threshold)
# accelerate offload (doesn't work properly)
if shared.args.gpu_memory or torch.cuda.device_count() > 1 or (is_xpu_available() and torch.xpu.device_count() > 1):
if shared.args.gpu_memory:
memory_map = list(map(lambda x: x.strip(), shared.args.gpu_memory))
max_cpu_memory = shared.args.cpu_memory.strip() if shared.args.cpu_memory is not None else '99GiB'
max_memory = {}
for i in range(len(memory_map)):
max_memory[i] = f'{memory_map[i]}GiB' if not re.match('.*ib$', memory_map[i].lower()) else memory_map[i]
max_memory['cpu'] = f'{max_cpu_memory}GiB' if not re.match('.*ib$', max_cpu_memory.lower()) else max_cpu_memory
else:
max_memory = accelerate.utils.get_balanced_memory(model)
device_map = accelerate.infer_auto_device_map(model, max_memory=max_memory, no_split_module_classes=["LlamaDecoderLayer"])
logger.info("Using the following device map for the quantized model:", device_map)
# https://huggingface.co/docs/accelerate/package_reference/big_modeling#accelerate.dispatch_model
model = accelerate.dispatch_model(model, device_map=device_map, offload_buffers=True)
# No offload
elif not shared.args.cpu:
if is_xpu_available():
model = model.to(torch.device("xpu:0"))
else:
model = model.to(torch.device('cuda:0'))
return model

8
modules/chat.py
View File

@ -308,9 +308,6 @@ def chatbot_wrapper(text, state, regenerate=False, _continue=False, loading_mess
'internal': output['internal']
}
if shared.model_name == 'None' or shared.model is None:
raise ValueError("No model is loaded! Select one in the Model tab.")
# Generate the prompt
kwargs = {
'_continue': _continue,
@ -355,11 +352,6 @@ def impersonate_wrapper(text, state):
static_output = chat_html_wrapper(state['history'], state['name1'], state['name2'], state['mode'], state['chat_style'], state['character_menu'])
if shared.model_name == 'None' or shared.model is None:
logger.error("No model is loaded! Select one in the Model tab.")
yield '', static_output
return
prompt = generate_chat_prompt('', state, impersonate=True)
stopping_strings = get_stopping_strings(state)

53
modules/grammar/grammar_utils.py
View File

@ -60,7 +60,7 @@ def hex_to_int(c):
return int(c)
elif "a" <= c.lower() <= "f":
return ord(c.lower()) - ord("a") + 10
return -1
raise RuntimeError("unknown hex char " + c)
def remove_leading_white_space(src, newline_ok):
@ -100,6 +100,13 @@ def parse_name(src):
return src[:pos], src[pos:]
def read_hex(s):
val = 0
for c in s:
val = (val << 4) + hex_to_int(c)
return chr(val)
def parse_char(src):
"""
parse the leading char from the input string
@ -111,13 +118,12 @@ def parse_char(src):
if src[0] == "\\":
esc = src[1]
if esc == "x":
first = hex_to_int(src[2])
if first > -1:
second = hex_to_int(src[3])
if second > -1:
return (first << 4) + second, src[4:]
raise RuntimeError("expecting \\xNN at " + src)
elif esc in ('"', "[", "]"):
return read_hex(src[2:4]), src[4:]
elif esc == "u":
return read_hex(src[2:6]), src[6:]
elif esc == "U":
return read_hex(src[2:10]), src[10:]
elif esc in ('"', "[", "]", "\\", "-"):
return esc, src[2:]
elif esc == "r":
return "\r", src[2:]
@ -125,6 +131,8 @@ def parse_char(src):
return "\n", src[2:]
elif esc == "t":
return "\t", src[2:]
elif esc == "\\":
return "\\", src[2:]
raise RuntimeError("unknown escape at " + src)
elif src:
return src[0], src[1:]
@ -454,7 +462,8 @@ class IncrementalGrammarConstraint(GrammarConstraint):
def __init__(self, grammar_str, start_rule_name, tokenizer):
super().__init__(grammar_str, start_rule_name, tokenizer)
def accept_char(self, byte, stacks):
def accept_char(self, char, stacks):
byte = ord(char)
new_stacks = []
for stack in stacks:
# stack is empty
@ -471,6 +480,9 @@ class IncrementalGrammarConstraint(GrammarConstraint):
if self.grammar_encoding[pos + i] <= byte and byte <= self.grammar_encoding[pos + i + 1]:
found = True
break
if self.grammar_encoding[pos + i] >= byte and byte >= self.grammar_encoding[pos + i + 1]:
found = True
break
if not found:
continue
@ -483,9 +495,8 @@ class IncrementalGrammarConstraint(GrammarConstraint):
return new_stacks
def accept_string(self, string: str, stacks: List[List[int]]):
_bytes = bytes(string, "utf-8")
for byte in _bytes:
stacks = self.accept_char(byte, stacks)
for char in string:
stacks = self.accept_char(char, stacks)
return stacks
def accept_token_id(self, token_id: int, stacks: List[List[int]]):
@ -537,16 +548,18 @@ class IncrementalGrammarConstraint(GrammarConstraint):
# For each sub-rule in the grammar, cache whether each byte is accepted.
@lru_cache(maxsize=None)
def pos_char_acceptance(self, pos):
acceptance = [False] * 256
def pos_char_acceptance(self, pos, char):
byte = ord(char)
num_chars = self.grammar_encoding[pos]
pos += 1
for i in range(0, num_chars, 2):
start = self.grammar_encoding[pos + i]
end = self.grammar_encoding[pos + i + 1]
for j in range(start, end + 1):
acceptance[j] = True
return acceptance
if byte >= start and byte <= end:
return True
if byte <= start and byte >= end:
return True
return False
# Probably this should be configurable. If the grammar has an exceedingly
# large number of states, the correct setting is a tradeoff between GPU
@ -580,7 +593,7 @@ class IncrementalGrammarConstraint(GrammarConstraint):
pos = stk[-1]
num_chars = self.grammar_encoding[pos]
if not self.pos_char_acceptance(pos)[byte]:
if not self.pos_char_acceptance(pos, byte):
continue
pos += num_chars + 1
@ -657,14 +670,14 @@ class TokenTrie:
token = tokenizer.convert_ids_to_tokens(id)
token = re.sub(r"<0x([0-9a-fA-F]{2})>", replace_hex, token)
token = token.replace("▁", " ")
return bytes(token, "utf-8")
return token
else:
print("Warning: unrecognized tokenizer: using default token formatting")
def fmt_token(id):
token = tokenizer.convert_ids_to_tokens(id)
return bytes(token, "utf-8")
return token
# note: vocab_size doesn't work here because there are also
# get_added_vocab() tokens

50
modules/loaders.py
View File

@ -46,7 +46,7 @@ loaders_and_params = OrderedDict({
'no_offload_kqv',
'row_split',
'tensorcores',
'flash-attn',
'flash_attn',
'streaming_llm',
'attention_sink_size',
],
@ -72,7 +72,7 @@ loaders_and_params = OrderedDict({
'no_offload_kqv',
'row_split',
'tensorcores',
'flash-attn',
'flash_attn',
'streaming_llm',
'attention_sink_size',
'llamacpp_HF_info',
@ -105,7 +105,6 @@ loaders_and_params = OrderedDict({
],
'AutoGPTQ': [
'triton',
'no_inject_fused_attention',
'no_inject_fused_mlp',
'no_use_cuda_fp16',
'wbits',
@ -131,21 +130,6 @@ loaders_and_params = OrderedDict({
'trust_remote_code',
'no_use_fast',
],
'GPTQ-for-LLaMa': [
'wbits',
'groupsize',
'model_type',
'pre_layer',
'trust_remote_code',
'no_use_fast',
'gptq_for_llama_info',
],
'QuIP#': [
'trust_remote_code',
'no_use_fast',
'no_flash_attn',
'quipsharp_info',
],
'HQQ': [
'hqq_backend',
'trust_remote_code',
@ -178,6 +162,10 @@ def transformers_samplers():
'repetition_penalty_range',
'encoder_repetition_penalty',
'no_repeat_ngram_size',
'dry_multiplier',
'dry_base',
'dry_allowed_length',
'dry_sequence_breakers',
'seed',
'do_sample',
'penalty_alpha',
@ -201,9 +189,7 @@ def transformers_samplers():
loaders_samplers = {
'Transformers': transformers_samplers(),
'AutoGPTQ': transformers_samplers(),
'GPTQ-for-LLaMa': transformers_samplers(),
'AutoAWQ': transformers_samplers(),
'QuIP#': transformers_samplers(),
'HQQ': transformers_samplers(),
'ExLlamav2': {
'temperature',
@ -251,6 +237,10 @@ loaders_samplers = {
'repetition_penalty_range',
'encoder_repetition_penalty',
'no_repeat_ngram_size',
'dry_multiplier',
'dry_base',
'dry_allowed_length',
'dry_sequence_breakers',
'seed',
'do_sample',
'mirostat_mode',
@ -309,6 +299,10 @@ loaders_samplers = {
'repetition_penalty_range',
'encoder_repetition_penalty',
'no_repeat_ngram_size',
'dry_multiplier',
'dry_base',
'dry_allowed_length',
'dry_sequence_breakers',
'seed',
'do_sample',
'mirostat_mode',
@ -327,15 +321,6 @@ loaders_samplers = {
},
}
loaders_model_types = {
'GPTQ-for-LLaMa': [
"None",
"llama",
"opt",
"gptj"
],
}
@functools.cache
def list_all_samplers():
@ -363,13 +348,6 @@ def blacklist_samplers(loader, dynamic_temperature):
return output
def get_model_types(loader):
if loader in loaders_model_types:
return loaders_model_types[loader]
return ["None"]
def get_gpu_memory_keys():
return [k for k in shared.gradio if k.startswith('gpu_memory')]

33
modules/logits.py
View File

@ -1,14 +1,34 @@
import time
import traceback
import torch
from transformers import is_torch_npu_available, is_torch_xpu_available
from modules import sampler_hijack, shared
from modules import models, sampler_hijack, shared
from modules.logging_colors import logger
from modules.models import load_model
from modules.text_generation import generate_reply
global_scores = None
def get_next_logits(prompt, state, use_samplers, previous, top_logits=25, return_dict=False):
def get_next_logits(*args, **kwargs):
if shared.args.idle_timeout > 0 and shared.model is None and shared.previous_model_name not in [None, 'None']:
shared.model, shared.tokenizer = load_model(shared.previous_model_name)
shared.generation_lock.acquire()
try:
result = _get_next_logits(*args, **kwargs)
except Exception:
traceback.print_exc()
result = None
models.last_generation_time = time.time()
shared.generation_lock.release()
return result
def _get_next_logits(prompt, state, use_samplers, previous, top_logits=25, return_dict=False):
if shared.model is None:
logger.error("No model is loaded! Select one in the Model tab.")
return 'Error: No model is loaded1 Select one in the Model tab.', previous
@ -66,7 +86,14 @@ def get_next_logits(prompt, state, use_samplers, previous, top_logits=25, return
topk_values = [float(i) for i in topk_values]
output = {}
for row in list(zip(topk_values, tokens)):
output[row[1]] = row[0]
key = row[1]
if isinstance(key, bytes):
try:
key = key.decode()
except:
key = key.decode('latin')
output[key] = row[0]
return output
else:

81
modules/models.py
View File

@ -1,5 +1,4 @@
import gc
import logging
import os
import pprint
import re
@ -29,7 +28,6 @@ import modules.shared as shared
from modules import RoPE, sampler_hijack
from modules.logging_colors import logger
from modules.models_settings import get_model_metadata
from modules.relative_imports import RelativeImport
transformers.logging.set_verbosity_error()
@ -61,6 +59,9 @@ if shared.args.deepspeed:
sampler_hijack.hijack_samplers()
last_generation_time = time.time()
def load_model(model_name, loader=None):
logger.info(f"Loading \"{model_name}\"")
t0 = time.time()
@ -70,13 +71,11 @@ def load_model(model_name, loader=None):
load_func_map = {
'Transformers': huggingface_loader,
'AutoGPTQ': AutoGPTQ_loader,
'GPTQ-for-LLaMa': GPTQ_loader,
'llama.cpp': llamacpp_loader,
'llamacpp_HF': llamacpp_HF_loader,
'ExLlamav2': ExLlamav2_loader,
'ExLlamav2_HF': ExLlamav2_HF_loader,
'AutoAWQ': AutoAWQ_loader,
'QuIP#': QuipSharp_loader,
'HQQ': HQQ_loader,
}
@ -266,7 +265,7 @@ def llamacpp_loader(model_name):
if path.is_file():
model_file = path
else:
model_file = list(Path(f'{shared.args.model_dir}/{model_name}').glob('*.gguf'))[0]
model_file = sorted(Path(f'{shared.args.model_dir}/{model_name}').glob('*.gguf'))[0]
logger.info(f"llama.cpp weights detected: \"{model_file}\"")
model, tokenizer = LlamaCppModel.from_pretrained(model_file)
@ -307,55 +306,6 @@ def AutoAWQ_loader(model_name):
return model
def QuipSharp_loader(model_name):
try:
with RelativeImport("repositories/quip-sharp"):
from lib.utils.unsafe_import import model_from_hf_path
except:
logger.error(
"\nQuIP# has not been found. It must be installed manually for now.\n"
"For instructions on how to do that, please consult:\n"
"https://github.com/oobabooga/text-generation-webui/pull/4803\n"
)
return None, None
# This fixes duplicate logging messages after the import above.
handlers = logging.getLogger().handlers
if len(handlers) > 1:
logging.getLogger().removeHandler(handlers[1])
model_dir = Path(f'{shared.args.model_dir}/{model_name}')
if not all((model_dir / file).exists() for file in ['tokenizer_config.json', 'special_tokens_map.json', 'tokenizer.model']):
logger.error(f"Could not load the model because the tokenizer files could not be found in the model folder. Please download the following files from the original (unquantized) model into {model_dir}: special_tokens_map.json, tokenizer.json, tokenizer.model, tokenizer_config.json.")
return None, None
model, model_str = model_from_hf_path(
model_dir,
use_cuda_graph=False,
use_flash_attn=not shared.args.no_flash_attn
)
return model
def GPTQ_loader(model_name):
# Monkey patch
if shared.args.monkey_patch:
logger.warning("Applying the monkey patch for using LoRAs with GPTQ models. It may cause undefined behavior outside its intended scope.")
from modules.monkey_patch_gptq_lora import load_model_llama
model, _ = load_model_llama(model_name)
# No monkey patch
else:
import modules.GPTQ_loader
model = modules.GPTQ_loader.load_quantized(model_name)
return model
def AutoGPTQ_loader(model_name):
import modules.AutoGPTQ_loader
@ -377,12 +327,12 @@ def ExLlamav2_HF_loader(model_name):
def HQQ_loader(model_name):
from hqq.core.quantize import HQQBackend, HQQLinear
from hqq.engine.hf import HQQModelForCausalLM
from hqq.models.hf.base import AutoHQQHFModel
logger.info(f"Loading HQQ model with backend: \"{shared.args.hqq_backend}\"")
model_dir = Path(f'{shared.args.model_dir}/{model_name}')
model = HQQModelForCausalLM.from_quantized(str(model_dir))
model = AutoHQQHFModel.from_quantized(str(model_dir))
HQQLinear.set_backend(getattr(HQQBackend, shared.args.hqq_backend))
return model
@ -428,6 +378,7 @@ def clear_torch_cache():
def unload_model():
shared.model = shared.tokenizer = None
shared.previous_model_name = shared.model_name
shared.model_name = 'None'
shared.lora_names = []
shared.model_dirty_from_training = False
@ -437,3 +388,21 @@ def unload_model():
def reload_model():
unload_model()
shared.model, shared.tokenizer = load_model(shared.model_name)
def unload_model_if_idle():
global last_generation_time
logger.info(f"Setting a timeout of {shared.args.idle_timeout} minutes to unload the model in case of inactivity.")
while True:
shared.generation_lock.acquire()
try:
if time.time() - last_generation_time > shared.args.idle_timeout * 60:
if shared.model is not None:
logger.info("Unloading the model for inactivity.")
unload_model()
finally:
shared.generation_lock.release()
time.sleep(60)

17
modules/models_settings.py
View File

@ -12,7 +12,6 @@ def get_fallback_settings():
'wbits': 'None',
'groupsize': 'None',
'desc_act': False,
'model_type': 'None',
'max_seq_len': 2048,
'n_ctx': 2048,
'rope_freq_base': 0,
@ -40,12 +39,7 @@ def get_model_metadata(model):
hf_metadata = None
if 'loader' not in model_settings:
if hf_metadata is not None and 'quip_params' in hf_metadata:
loader = 'QuIP#'
else:
loader = infer_loader(model, model_settings)
model_settings['loader'] = loader
model_settings['loader'] = infer_loader(model, model_settings)
# GGUF metadata
if model_settings['loader'] in ['llama.cpp', 'llamacpp_HF']:
@ -204,20 +198,17 @@ def update_model_parameters(state, initial=False):
continue
# Setting null defaults
if element in ['wbits', 'groupsize', 'model_type'] and value == 'None':
if element in ['wbits', 'groupsize'] and value == 'None':
value = vars(shared.args_defaults)[element]
elif element in ['cpu_memory'] and value == 0:
value = vars(shared.args_defaults)[element]
# Making some simple conversions
if element in ['wbits', 'groupsize', 'pre_layer']:
if element in ['wbits', 'groupsize']:
value = int(value)
elif element == 'cpu_memory' and value is not None:
value = f"{value}MiB"
if element in ['pre_layer']:
value = [value] if value > 0 else None
setattr(shared.args, element, value)
found_positive = False
@ -242,7 +233,7 @@ def apply_model_settings_to_state(model, state):
loader = model_settings.pop('loader')
# If the user is using an alternative loader for the same model type, let them keep using it
if not (loader == 'ExLlamav2_HF' and state['loader'] in ['GPTQ-for-LLaMa', 'ExLlamav2', 'AutoGPTQ']):
if not (loader == 'ExLlamav2_HF' and state['loader'] in ['ExLlamav2', 'AutoGPTQ']):
state['loader'] = loader
for k in model_settings:

39
modules/monkey_patch_gptq_lora.py
View File

@ -1,39 +0,0 @@
# Copied from https://github.com/johnsmith0031/alpaca_lora_4bit
from pathlib import Path
import alpaca_lora_4bit.autograd_4bit as autograd_4bit
from alpaca_lora_4bit.amp_wrapper import AMPWrapper
from alpaca_lora_4bit.autograd_4bit import (
Autograd4bitQuantLinear,
load_llama_model_4bit_low_ram
)
from alpaca_lora_4bit.models import Linear4bitLt
from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
replace_peft_model_with_int4_lora_model
)
from modules import shared
from modules.GPTQ_loader import find_quantized_model_file
replace_peft_model_with_int4_lora_model()
def load_model_llama(model_name):
config_path = str(Path(f'{shared.args.model_dir}/{model_name}'))
model_path = str(find_quantized_model_file(model_name))
model, tokenizer = load_llama_model_4bit_low_ram(config_path, model_path, groupsize=shared.args.groupsize, is_v1_model=False)
for _, m in model.named_modules():
if isinstance(m, Autograd4bitQuantLinear) or isinstance(m, Linear4bitLt):
if m.is_v1_model:
m.zeros = m.zeros.half()
m.scales = m.scales.half()
m.bias = m.bias.half()
autograd_4bit.auto_switch = True
model.half()
wrapper = AMPWrapper(model)
wrapper.apply_generate()
return model, tokenizer

4
modules/presets.py
View File

@ -40,6 +40,10 @@ def default_preset():
'do_sample': True,
'encoder_repetition_penalty': 1,
'no_repeat_ngram_size': 0,
'dry_multiplier': 0,
'dry_base': 1.75,
'dry_allowed_length': 2,
'dry_sequence_breakers': '"\\n", ":", "\\"", "*"',
'sampler_priority': 'temperature\ndynamic_temperature\nquadratic_sampling\ntop_k\ntop_p\ntypical_p\nepsilon_cutoff\neta_cutoff\ntfs\ntop_a\nmin_p\nmirostat'
}

153
modules/sampler_hijack.py
View File

@ -1,3 +1,4 @@
import json
import math
import pprint
@ -124,36 +125,6 @@ class QuadraticSamplingLogitsWarper(LogitsWarper):
return transformed_logits
class MinPLogitsWarper(LogitsWarper):
def __init__(self, min_p: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
if min_p < 0 or min_p > 1.0:
raise ValueError(f"`min_p` has to be a float >= 0 and <= 1, but is {min_p}")
self.min_p = min_p
self.filter_value = filter_value
self.min_tokens_to_keep = min_tokens_to_keep
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
# Convert logits to probabilities
probs = torch.softmax(scores, dim=-1)
# Get the probability of the top token for each sequence in the batch
top_probs, _ = probs.max(dim=-1, keepdim=True)
# Calculate the actual min_p threshold by scaling min_p with the top token's probability
scaled_min_p = self.min_p * top_probs
# Create a mask for tokens that have a probability less than the scaled min_p
tokens_to_remove = probs < scaled_min_p
sorted_indices = torch.argsort(scores, descending=True, dim=-1)
sorted_indices_to_remove = torch.gather(tokens_to_remove, dim=-1, index=sorted_indices)
if self.min_tokens_to_keep > 1:
# Keep at least min_tokens_to_keep
sorted_indices_to_remove[..., : self.min_tokens_to_keep] = False
indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
scores = scores.masked_fill(indices_to_remove, self.filter_value)
return scores
class TailFreeLogitsWarper(LogitsWarper):
def __init__(self, tfs: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
tfs = float(tfs)
@ -220,6 +191,78 @@ class TopALogitsWarper(LogitsWarper):
return scores
class DRYLogitsProcessor(LogitsProcessor):
def __init__(self, multiplier: float, base: float, allowed_length: int, sequence_breakers: set[int], _range: int):
self.multiplier = multiplier
self.base = base
self.allowed_length = allowed_length
self.sequence_breakers = sequence_breakers
self._range = _range
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
if self._range > 0:
input_ids = input_ids[:, -self._range:]
for input_ids_row, scores_row in zip(input_ids, scores):
# Use normal Python data types for improved performance
input_ids = input_ids_row.tolist()
last_token = input_ids[-1]
if last_token in self.sequence_breakers:
continue
# Exclude the last token as it always matches.
match_indices = []
for idx, val in enumerate(input_ids[:-1]):
if val == last_token:
match_indices.append(idx)
# Stores the maximum matching sequence length
# for each token immediately following the sequence in the input.
match_lengths = {}
for i in match_indices:
next_token = input_ids[i + 1]
if next_token in self.sequence_breakers:
continue
# We have already found that `last_token` matches at this index,
# so the match is at least of length 1.
match_length = 1
# Extend the match backwards (at most to 50 to prevent exponent overflow at penalty calculation) (this cap also improves performance on worst case)
while match_length < 50:
j = i - match_length
if j < 0:
# Start of input reached.
break
previous_token = input_ids[-(match_length + 1)]
if input_ids[j] != previous_token:
# Start of match reached.
break
if previous_token in self.sequence_breakers:
# Sequence-breaking token reached.
break
match_length += 1
if next_token in match_lengths:
match_lengths[next_token] = max(match_length, match_lengths[next_token])
else:
match_lengths[next_token] = match_length
# Apply penalties.
for token, match_length in match_lengths.items():
if match_length >= self.allowed_length:
penalty = self.multiplier * self.base ** (match_length - self.allowed_length)
scores_row[token] -= penalty
return scores
class MirostatLogitsWarper(LogitsWarper):
def __init__(self, mirostat_mode: int, mirostat_tau: float, mirostat_eta: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
if mirostat_mode not in [2]:
@ -352,14 +395,6 @@ def get_logits_warper_patch(self, generation_config):
)
)
if generation_config.min_p is not None and 0.0 < generation_config.min_p <= 1.0:
warpers_to_add.append(
MinPLogitsWarper(
min_p=generation_config.min_p,
min_tokens_to_keep=min_tokens_to_keep
)
)
if generation_config.dynamic_temperature:
warpers_to_add.append(
DynamicTemperatureLogitsWarper(
@ -448,20 +483,44 @@ def get_logits_warper_patch(self, generation_config):
def get_logits_processor_patch(self, **kwargs):
repetition_penalty = kwargs['generation_config'].repetition_penalty
presence_penalty = kwargs['generation_config'].presence_penalty
frequency_penalty = kwargs['generation_config'].frequency_penalty
repetition_penalty_range = kwargs['generation_config'].repetition_penalty_range
do_rep_pen_hijack = (repetition_penalty > 1) or (presence_penalty != 0) or (frequency_penalty != 0)
generation_config = kwargs['generation_config']
do_rep_pen_hijack = (generation_config.repetition_penalty > 1) or (generation_config.presence_penalty != 0) or (generation_config.frequency_penalty != 0)
if do_rep_pen_hijack:
kwargs['generation_config'].repetition_penalty = 1.1 # Set to value > 1 to ensure RepetitionPenaltyLogitsProcessor is created
generation_config.repetition_penalty = 1.1 # Set to value > 1 to ensure RepetitionPenaltyLogitsProcessor is created
result = self._get_logits_processor_old(**kwargs)
if do_rep_pen_hijack:
for i in range(len(result)):
if result[i].__class__.__name__ == 'RepetitionPenaltyLogitsProcessor':
result[i] = RepetitionPenaltyLogitsProcessorWithRange(repetition_penalty, presence_penalty, frequency_penalty, repetition_penalty_range)
result[i] = RepetitionPenaltyLogitsProcessorWithRange(
generation_config.repetition_penalty,
generation_config.presence_penalty,
generation_config.frequency_penalty,
generation_config.repetition_penalty_range
)
if generation_config.dry_multiplier is not None and generation_config.dry_multiplier > 0.0:
dry_sequence_breakers = generation_config.dry_sequence_breakers
# Support both JSON array notation and comma-separated strings.
if not dry_sequence_breakers.startswith("["):
dry_sequence_breakers = "[" + dry_sequence_breakers + "]"
sequence_breaker_strings = json.loads(dry_sequence_breakers)
# Prefix with 'a' to get the correct encoding of the token at the end of a text.
sequence_breakers = {shared.tokenizer.encode(f'a{s}')[-1] for s in sequence_breaker_strings}
result.append(
DRYLogitsProcessor(
multiplier=generation_config.dry_multiplier,
base=generation_config.dry_base,
allowed_length=generation_config.dry_allowed_length,
sequence_breakers=sequence_breakers,
_range=generation_config.repetition_penalty_range,
)
)
return result
@ -483,6 +542,10 @@ def generation_config_init_patch(self, **kwargs):
self.repetition_penalty_range = kwargs.pop("repetition_penalty_range", 0)
self.presence_penalty = kwargs.pop("presence_penalty", 0)
self.frequency_penalty = kwargs.pop("frequency_penalty", 0)
self.dry_multiplier = kwargs.pop("dry_multiplier", 0.0)
self.dry_base = kwargs.pop("dry_base", 1.75)
self.dry_allowed_length = kwargs.pop("dry_allowed_length", 2)
self.dry_sequence_breakers = kwargs.pop("dry_sequence_breakers", '"\\n", ":", "\\"", "*"')
self.temperature_last = kwargs.pop("temperature_last", False)
self.sampler_priority = kwargs.pop("sampler_priority", ['temperature', 'dynamic_temperature', 'quadratic_sampling', 'top_k', 'top_p', 'typical_p', 'epsilon_cutoff', 'eta_cutoff', 'tfs', 'top_a', 'min_p', 'mirostat'])

26
modules/shared.py
View File

@ -13,6 +13,7 @@ from modules.logging_colors import logger
model = None
tokenizer = None
model_name = 'None'
previous_model_name = 'None'
is_seq2seq = False
model_dirty_from_training = False
lora_names = []
@ -84,10 +85,11 @@ group.add_argument('--settings', type=str, help='Load the default interface sett
group.add_argument('--extensions', type=str, nargs='+', help='The list of extensions to load. If you want to load more than one extension, write the names separated by spaces.')
group.add_argument('--verbose', action='store_true', help='Print the prompts to the terminal.')
group.add_argument('--chat-buttons', action='store_true', help='Show buttons on the chat tab instead of a hover menu.')
group.add_argument('--idle-timeout', type=int, default=0, help='Unload model after this many minutes of inactivity. It will be automatically reloaded when you try to use it again.')
# Model loader
group = parser.add_argument_group('Model loader')
group.add_argument('--loader', type=str, help='Choose the model loader manually, otherwise, it will get autodetected. Valid options: Transformers, llama.cpp, llamacpp_HF, ExLlamav2_HF, ExLlamav2, AutoGPTQ, AutoAWQ, GPTQ-for-LLaMa, QuIP#.')
group.add_argument('--loader', type=str, help='Choose the model loader manually, otherwise, it will get autodetected. Valid options: Transformers, llama.cpp, llamacpp_HF, ExLlamav2_HF, ExLlamav2, AutoGPTQ, AutoAWQ.')
# Transformers/Accelerate
group = parser.add_argument_group('Transformers/Accelerate')
@ -147,21 +149,17 @@ group.add_argument('--num_experts_per_token', type=int, default=2, help='Number
# AutoGPTQ
group = parser.add_argument_group('AutoGPTQ')
group.add_argument('--triton', action='store_true', help='Use triton.')
group.add_argument('--no_inject_fused_attention', action='store_true', help='Disable the use of fused attention, which will use less VRAM at the cost of slower inference.')
group.add_argument('--no_inject_fused_mlp', action='store_true', help='Triton mode only: disable the use of fused MLP, which will use less VRAM at the cost of slower inference.')
group.add_argument('--no_use_cuda_fp16', action='store_true', help='This can make models faster on some systems.')
group.add_argument('--desc_act', action='store_true', help='For models that do not have a quantize_config.json, this parameter is used to define whether to set desc_act or not in BaseQuantizeConfig.')
group.add_argument('--disable_exllama', action='store_true', help='Disable ExLlama kernel, which can improve inference speed on some systems.')
group.add_argument('--disable_exllamav2', action='store_true', help='Disable ExLlamav2 kernel.')
# GPTQ-for-LLaMa
group = parser.add_argument_group('GPTQ-for-LLaMa')
group.add_argument('--wbits', type=int, default=0, help='Load a pre-quantized model with specified precision in bits. 2, 3, 4 and 8 are supported.')
group.add_argument('--model_type', type=str, help='Model type of pre-quantized model. Currently LLaMA, OPT, and GPT-J are supported.')
group.add_argument('--groupsize', type=int, default=-1, help='Group size.')
group.add_argument('--pre_layer', type=int, nargs='+', help='The number of layers to allocate to the GPU. Setting this parameter enables CPU offloading for 4-bit models. For multi-gpu, write the numbers separated by spaces, eg --pre_layer 30 60.')
group.add_argument('--checkpoint', type=str, help='The path to the quantized checkpoint file. If not specified, it will be automatically detected.')
group.add_argument('--monkey-patch', action='store_true', help='Apply the monkey patch for using LoRAs with quantized models.')
# AutoAWQ
group = parser.add_argument_group('AutoAWQ')
group.add_argument('--no_inject_fused_attention', action='store_true', help='Disable the use of fused attention, which will use less VRAM at the cost of slower inference.')
# HQQ
group = parser.add_argument_group('HQQ')
@ -206,7 +204,11 @@ group = parser.add_argument_group('Multimodal')
group.add_argument('--multimodal-pipeline', type=str, default=None, help='The multimodal pipeline to use. Examples: llava-7b, llava-13b.')
# Deprecated parameters
# group = parser.add_argument_group('Deprecated')
group = parser.add_argument_group('Deprecated')
group.add_argument('--model_type', type=str, help='DEPRECATED')
group.add_argument('--pre_layer', type=int, nargs='+', help='DEPRECATED')
group.add_argument('--checkpoint', type=str, help='DEPRECATED')
group.add_argument('--monkey-patch', action='store_true', help='DEPRECATED')
args = parser.parse_args()
args_defaults = parser.parse_args([])
@ -251,8 +253,6 @@ def fix_loader_name(name):
return 'Transformers'
elif name in ['autogptq', 'auto-gptq', 'auto_gptq', 'auto gptq']:
return 'AutoGPTQ'
elif name in ['gptq-for-llama', 'gptqforllama', 'gptqllama', 'gptq for llama', 'gptq_for_llama']:
return 'GPTQ-for-LLaMa'
elif name in ['exllama', 'ex-llama', 'ex_llama', 'exlama']:
return 'ExLlama'
elif name in ['exllamav2', 'exllama-v2', 'ex_llama-v2', 'exlamav2', 'exlama-v2', 'exllama2', 'exllama-2']:
@ -261,8 +261,6 @@ def fix_loader_name(name):
return 'ExLlamav2_HF'
elif name in ['autoawq', 'awq', 'auto-awq']:
return 'AutoAWQ'
elif name in ['quip#', 'quip-sharp', 'quipsharp', 'quip_sharp']:
return 'QuIP#'
elif name in ['hqq']:
return 'HQQ'

27
modules/text_generation.py
View File

@ -16,6 +16,7 @@ from transformers import (
)
import modules.shared as shared
from modules import models
from modules.cache_utils import process_llamacpp_cache
from modules.callbacks import (
Iteratorize,
@ -27,15 +28,19 @@ from modules.grammar.grammar_utils import initialize_grammar
from modules.grammar.logits_process import GrammarConstrainedLogitsProcessor
from modules.html_generator import generate_basic_html
from modules.logging_colors import logger
from modules.models import clear_torch_cache
from modules.models import clear_torch_cache, load_model
def generate_reply(*args, **kwargs):
if shared.args.idle_timeout > 0 and shared.model is None and shared.previous_model_name not in [None, 'None']:
shared.model, shared.tokenizer = load_model(shared.previous_model_name)
shared.generation_lock.acquire()
try:
for result in _generate_reply(*args, **kwargs):
yield result
finally:
models.last_generation_time = time.time()
shared.generation_lock.release()
@ -133,9 +138,21 @@ def encode(prompt, add_special_tokens=True, add_bos_token=True, truncation_lengt
input_ids = np.array(input_ids).reshape(1, len(input_ids))
else:
input_ids = shared.tokenizer.encode(str(prompt), return_tensors='pt', add_special_tokens=add_special_tokens)
if not add_bos_token:
while len(input_ids[0]) > 0 and input_ids[0][0] == shared.tokenizer.bos_token_id:
input_ids = input_ids[:, 1:]
if hasattr(shared.tokenizer, 'bos_token_id') and shared.tokenizer.bos_token_id is not None:
if add_bos_token:
if (len(input_ids[0]) > 0 and input_ids[0][0] != shared.tokenizer.bos_token_id) or len(input_ids[0]) == 0:
# Add a missing bos token (it may not have been added due to faulty model metadata)
bos_tensor = torch.tensor([[shared.tokenizer.bos_token_id]])
input_ids = torch.cat((bos_tensor, input_ids), 1)
# Prevent double bos token due to jinja templates with <s> somewhere
while len(input_ids[0]) > 1 and input_ids[0][0] == shared.tokenizer.bos_token_id and input_ids[0][1] == shared.tokenizer.bos_token_id:
input_ids = input_ids[:, 1:]
else:
# Remove any bos token that may have been added
while len(input_ids[0]) > 0 and input_ids[0][0] == shared.tokenizer.bos_token_id:
input_ids = input_ids[:, 1:]
# Handling truncation
if truncation_length is not None:
@ -267,7 +284,7 @@ def get_reply_from_output_ids(output_ids, state=None, starting_from=0):
def generate_reply_HF(question, original_question, seed, state, stopping_strings=None, is_chat=False):
generate_params = {}
for k in ['max_new_tokens', 'temperature', 'temperature_last', 'dynamic_temperature', 'dynatemp_low', 'dynatemp_high', 'dynatemp_exponent', 'smoothing_factor', 'smoothing_curve', 'top_p', 'min_p', 'top_k', 'repetition_penalty', 'presence_penalty', 'frequency_penalty', 'repetition_penalty_range', 'typical_p', 'tfs', 'top_a', 'guidance_scale', 'penalty_alpha', 'mirostat_mode', 'mirostat_tau', 'mirostat_eta', 'do_sample', 'encoder_repetition_penalty', 'no_repeat_ngram_size']:
for k in ['max_new_tokens', 'temperature', 'temperature_last', 'dynamic_temperature', 'dynatemp_low', 'dynatemp_high', 'dynatemp_exponent', 'smoothing_factor', 'smoothing_curve', 'top_p', 'min_p', 'top_k', 'repetition_penalty', 'presence_penalty', 'frequency_penalty', 'repetition_penalty_range', 'typical_p', 'tfs', 'top_a', 'guidance_scale', 'penalty_alpha', 'mirostat_mode', 'mirostat_tau', 'mirostat_eta', 'do_sample', 'encoder_repetition_penalty', 'no_repeat_ngram_size', 'dry_multiplier', 'dry_base', 'dry_allowed_length', 'dry_sequence_breakers']:
if k in state:
generate_params[k] = state[k]

19
modules/training.py
View File

@ -292,12 +292,6 @@ def calc_trainable_parameters(model):
def do_train(lora_name: str, always_override: bool, q_proj_en: bool, v_proj_en: bool, k_proj_en: bool, o_proj_en: bool, gate_proj_en: bool, down_proj_en: bool, up_proj_en: bool, save_steps: int, micro_batch_size: int, batch_size: int, epochs: int, learning_rate: str, lr_scheduler_type: str, lora_rank: int, lora_alpha: int, lora_dropout: float, cutoff_len: int, dataset: str, eval_dataset: str, format: str, eval_steps: int, raw_text_file: str, overlap_len: int, newline_favor_len: int, higher_rank_limit: bool, warmup_steps: int, optimizer: str, hard_cut_string: str, train_only_after: str, stop_at_loss: float, add_eos_token: bool, min_chars: int, report_to: str):
if shared.args.monkey_patch:
from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
replace_peft_model_with_int4_lora_model
)
replace_peft_model_with_int4_lora_model()
global WANT_INTERRUPT
WANT_INTERRUPT = False
@ -329,10 +323,6 @@ def do_train(lora_name: str, always_override: bool, q_proj_en: bool, v_proj_en:
time.sleep(5)
if shared.args.loader == 'GPTQ-for-LLaMa' and not shared.args.monkey_patch:
yield "LoRA training with GPTQ-for-LLaMa requires loading with `--monkey-patch`"
return
if cutoff_len <= 0 or micro_batch_size <= 0 or batch_size <= 0 or actual_lr <= 0 or lora_rank <= 0 or lora_alpha <= 0:
yield "Cannot input zeroes."
return
@ -553,15 +543,6 @@ def do_train(lora_name: str, always_override: bool, q_proj_en: bool, v_proj_en:
yield traceback.format_exc().replace('\n', '\n\n')
return
if shared.args.monkey_patch:
from alpaca_lora_4bit.autograd_4bit import Autograd4bitQuantLinear
from alpaca_lora_4bit.models import Linear4bitLt
for _, m in lora_model.named_modules():
if isinstance(m, Autograd4bitQuantLinear) or isinstance(m, Linear4bitLt):
if m.is_v1_model:
m.zeros = m.zeros.half()
m.scales = m.scales.half()
class Tracked():
def __init__(self):
self.current_steps = 0

8
modules/ui.py
View File

@ -70,8 +70,6 @@ def list_model_elements():
'use_double_quant',
'wbits',
'groupsize',
'model_type',
'pre_layer',
'triton',
'desc_act',
'no_inject_fused_attention',
@ -104,7 +102,7 @@ def list_model_elements():
'no_offload_kqv',
'row_split',
'tensorcores',
'flash-attn',
'flash_attn',
'streaming_llm',
'attention_sink_size',
'hqq_backend',
@ -147,6 +145,10 @@ def list_interface_input_elements():
'repetition_penalty_range',
'encoder_repetition_penalty',
'no_repeat_ngram_size',
'dry_multiplier',
'dry_base',
'dry_allowed_length',
'dry_sequence_breakers',
'do_sample',
'penalty_alpha',
'mirostat_mode',

25
modules/ui_model_menu.py
View File

@ -97,12 +97,10 @@ def create_ui():
shared.gradio['n_ctx'] = gr.Slider(minimum=0, maximum=shared.settings['truncation_length_max'], step=256, label="n_ctx", value=shared.args.n_ctx, info='Context length. Try lowering this if you run out of memory while loading the model.')
shared.gradio['tensor_split'] = gr.Textbox(label='tensor_split', info='List of proportions to split the model across multiple GPUs. Example: 18,17')
shared.gradio['n_batch'] = gr.Slider(label="n_batch", minimum=1, maximum=2048, step=1, value=shared.args.n_batch)
shared.gradio['threads'] = gr.Slider(label="threads", minimum=0, step=1, maximum=32, value=shared.args.threads)
shared.gradio['threads_batch'] = gr.Slider(label="threads_batch", minimum=0, step=1, maximum=32, value=shared.args.threads_batch)
shared.gradio['threads'] = gr.Slider(label="threads", minimum=0, step=1, maximum=256, value=shared.args.threads)
shared.gradio['threads_batch'] = gr.Slider(label="threads_batch", minimum=0, step=1, maximum=256, value=shared.args.threads_batch)
shared.gradio['wbits'] = gr.Dropdown(label="wbits", choices=["None", 1, 2, 3, 4, 8], value=shared.args.wbits if shared.args.wbits > 0 else "None")
shared.gradio['groupsize'] = gr.Dropdown(label="groupsize", choices=["None", 32, 64, 128, 1024], value=shared.args.groupsize if shared.args.groupsize > 0 else "None")
shared.gradio['model_type'] = gr.Dropdown(label="model_type", choices=["None"], value=shared.args.model_type or "None")
shared.gradio['pre_layer'] = gr.Slider(label="pre_layer", minimum=0, maximum=100, value=shared.args.pre_layer[0] if shared.args.pre_layer is not None else 0)
shared.gradio['gpu_split'] = gr.Textbox(label='gpu-split', info='Comma-separated list of VRAM (in GB) to use per GPU. Example: 20,7,7')
shared.gradio['max_seq_len'] = gr.Slider(label='max_seq_len', minimum=0, maximum=shared.settings['truncation_length_max'], step=256, info='Context length. Try lowering this if you run out of memory while loading the model.', value=shared.args.max_seq_len)
with gr.Blocks():
@ -111,14 +109,13 @@ def create_ui():
shared.gradio['compress_pos_emb'] = gr.Slider(label='compress_pos_emb', minimum=1, maximum=8, step=1, info='Positional embeddings compression factor. Should be set to (context length) / (model\'s original context length). Equal to 1/rope_freq_scale.', value=shared.args.compress_pos_emb)
shared.gradio['autogptq_info'] = gr.Markdown('ExLlamav2_HF is recommended over AutoGPTQ for models derived from Llama.')
shared.gradio['quipsharp_info'] = gr.Markdown('QuIP# has to be installed manually at the moment.')
with gr.Column():
shared.gradio['load_in_8bit'] = gr.Checkbox(label="load-in-8bit", value=shared.args.load_in_8bit)
shared.gradio['load_in_4bit'] = gr.Checkbox(label="load-in-4bit", value=shared.args.load_in_4bit)
shared.gradio['use_double_quant'] = gr.Checkbox(label="use_double_quant", value=shared.args.use_double_quant)
shared.gradio['use_flash_attention_2'] = gr.Checkbox(label="use_flash_attention_2", value=shared.args.use_flash_attention_2, info='Set use_flash_attention_2=True while loading the model.')
shared.gradio['flash-attn'] = gr.Checkbox(label="flash-attn", value=shared.args.flash_attn, info='Use flash-attention.')
shared.gradio['flash_attn'] = gr.Checkbox(label="flash_attn", value=shared.args.flash_attn, info='Use flash-attention.')
shared.gradio['auto_devices'] = gr.Checkbox(label="auto-devices", value=shared.args.auto_devices)
shared.gradio['tensorcores'] = gr.Checkbox(label="tensorcores", value=shared.args.tensorcores, info='NVIDIA only: use llama-cpp-python compiled with tensor cores support. This increases performance on RTX cards.')
shared.gradio['streaming_llm'] = gr.Checkbox(label="streaming_llm", value=shared.args.streaming_llm, info='(experimental) Activate StreamingLLM to avoid re-evaluating the entire prompt when old messages are removed.')
@ -187,9 +184,7 @@ def create_ui():
def create_event_handlers():
shared.gradio['loader'].change(
loaders.make_loader_params_visible, gradio('loader'), gradio(loaders.get_all_params())).then(
lambda value: gr.update(choices=loaders.get_model_types(value)), gradio('loader'), gradio('model_type'))
shared.gradio['loader'].change(loaders.make_loader_params_visible, gradio('loader'), gradio(loaders.get_all_params()))
# In this event handler, the interface state is read and updated
# with the model defaults (if any), and then the model is loaded
@ -273,6 +268,10 @@ def load_lora_wrapper(selected_loras):
def download_model_wrapper(repo_id, specific_file, progress=gr.Progress(), return_links=False, check=False):
try:
if repo_id == "":
yield ("Please enter a model path")
return
downloader = importlib.import_module("download-model").ModelDownloader()
progress(0.0)
@ -290,7 +289,13 @@ def download_model_wrapper(repo_id, specific_file, progress=gr.Progress(), retur
return
yield ("Getting the output folder")
output_folder = downloader.get_output_folder(model, branch, is_lora, is_llamacpp=is_llamacpp)
output_folder = downloader.get_output_folder(
model,
branch,
is_lora,
is_llamacpp=is_llamacpp,
model_dir=shared.args.model_dir if shared.args.model_dir != shared.args_defaults.model_dir else None
)
if output_folder == Path("models"):
output_folder = Path(shared.args.model_dir)

8
modules/ui_parameters.py
View File

@ -38,6 +38,14 @@ def create_ui(default_preset):
shared.gradio['presence_penalty'] = gr.Slider(0, 2, value=generate_params['presence_penalty'], step=0.05, label='presence_penalty')
shared.gradio['repetition_penalty_range'] = gr.Slider(0, 4096, step=64, value=generate_params['repetition_penalty_range'], label='repetition_penalty_range')
shared.gradio['do_sample'] = gr.Checkbox(value=generate_params['do_sample'], label='do_sample')
with gr.Blocks():
gr.Markdown("[DRY sequence repetition penalty](https://github.com/oobabooga/text-generation-webui/pull/5677)")
shared.gradio['dry_multiplier'] = gr.Slider(0, 5, value=generate_params['dry_multiplier'], step=0.01, label='dry_multiplier', info='Set to value > 0 to enable DRY. Controls the magnitude of the penalty for the shortest penalized sequences.')
shared.gradio['dry_base'] = gr.Slider(1, 4, value=generate_params['dry_base'], step=0.01, label='dry_base', info='Controls how fast the penalty grows with increasing sequence length.')
shared.gradio['dry_allowed_length'] = gr.Slider(1, 20, value=generate_params['dry_allowed_length'], step=1, label='dry_allowed_length', info='Longest sequence that can be repeated without being penalized.')
shared.gradio['dry_sequence_breakers'] = gr.Textbox(value=generate_params['dry_sequence_breakers'], label='dry_sequence_breakers', info='Tokens across which sequence matching is not continued. Specified as a comma-separated list of quoted strings.')
gr.Markdown("[Learn more](https://github.com/oobabooga/text-generation-webui/wiki/03-%E2%80%90-Parameters-Tab)")
with gr.Column():

2
one_click.py
View File

@ -388,7 +388,7 @@ def update_requirements(initial_installation=False, pull=True):
# Prepare the requirements file
textgen_requirements = open(requirements_file).read().splitlines()
if is_cuda118:
textgen_requirements = [req.replace('+cu121', '+cu118').replace('+cu122', '+cu118') for req in textgen_requirements]
textgen_requirements = [req.replace('+cu121', '+cu118').replace('+cu122', '+cu118') for req in textgen_requirements if "auto-gptq" not in req]
if is_windows() and is_cuda118: # No flash-attention on Windows for CUDA 11
textgen_requirements = [req for req in textgen_requirements if 'oobabooga/flash-attention' not in req]

19
requirements.txt
View File

@ -1,11 +1,12 @@
accelerate==0.27.*
aqlm[gpu,cpu]==1.1.3; platform_system == "Linux"
accelerate==0.30.*
aqlm[gpu,cpu]==1.1.5; platform_system == "Linux"
auto-gptq==0.7.1
bitsandbytes==0.43.*
colorama
datasets
einops
gradio==4.26.*
hqq==0.1.5
hqq==0.1.7.post2
jinja2==3.1.2
lm_eval==0.3.0
markdown
@ -23,7 +24,7 @@ safetensors==0.4.*
scipy
sentencepiece
tensorboard
transformers==4.40.*
transformers==4.41.*
tqdm
wandb
@ -52,10 +53,6 @@ https://github.com/oobabooga/llama-cpp-python-cuBLAS-wheels/releases/download/te
https://github.com/oobabooga/llama-cpp-python-cuBLAS-wheels/releases/download/textgen-webui/llama_cpp_python_cuda_tensorcores-0.2.75+cu121-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
# CUDA wheels
https://github.com/jllllll/AutoGPTQ/releases/download/v0.6.0/auto_gptq-0.6.0+cu121-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
https://github.com/jllllll/AutoGPTQ/releases/download/v0.6.0/auto_gptq-0.6.0+cu121-cp310-cp310-win_amd64.whl; platform_system == "Windows" and python_version == "3.10"
https://github.com/jllllll/AutoGPTQ/releases/download/v0.6.0/auto_gptq-0.6.0+cu121-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/jllllll/AutoGPTQ/releases/download/v0.6.0/auto_gptq-0.6.0+cu121-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
https://github.com/oobabooga/exllamav2/releases/download/v0.0.20/exllamav2-0.0.20+cu121-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
https://github.com/oobabooga/exllamav2/releases/download/v0.0.20/exllamav2-0.0.20+cu121-cp310-cp310-win_amd64.whl; platform_system == "Windows" and python_version == "3.10"
https://github.com/oobabooga/exllamav2/releases/download/v0.0.20/exllamav2-0.0.20+cu121-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
@ -65,8 +62,4 @@ https://github.com/oobabooga/flash-attention/releases/download/v2.5.6/flash_attn
https://github.com/oobabooga/flash-attention/releases/download/v2.5.6/flash_attn-2.5.6+cu122torch2.2.0cxx11abiFALSE-cp310-cp310-win_amd64.whl; platform_system == "Windows" and python_version == "3.10"
https://github.com/Dao-AILab/flash-attention/releases/download/v2.5.6/flash_attn-2.5.6+cu122torch2.2cxx11abiFALSE-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/Dao-AILab/flash-attention/releases/download/v2.5.6/flash_attn-2.5.6+cu122torch2.2cxx11abiFALSE-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+cu121-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+cu121-cp310-cp310-win_amd64.whl; platform_system == "Windows" and python_version == "3.10"
https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+cu121-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+cu121-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
autoawq==0.2.3; platform_system == "Linux" or platform_system == "Windows"
autoawq==0.2.5; platform_system == "Linux" or platform_system == "Windows"

14
requirements_amd.txt
View File

@ -1,9 +1,9 @@
accelerate==0.27.*
accelerate==0.30.*
colorama
datasets
einops
gradio==4.26.*
hqq==0.1.5
hqq==0.1.7.post2
jinja2==3.1.2
lm_eval==0.3.0
markdown
@ -21,7 +21,7 @@ safetensors==0.4.*
scipy
sentencepiece
tensorboard
transformers==4.40.*
transformers==4.41.*
tqdm
wandb
@ -40,12 +40,8 @@ https://github.com/oobabooga/llama-cpp-python-cuBLAS-wheels/releases/download/cp
# AMD wheels
https://github.com/oobabooga/llama-cpp-python-cuBLAS-wheels/releases/download/rocm/llama_cpp_python_cuda-0.2.75+rocm5.6.1-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/oobabooga/llama-cpp-python-cuBLAS-wheels/releases/download/rocm/llama_cpp_python_cuda-0.2.75+rocm5.6.1-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
https://github.com/jllllll/AutoGPTQ/releases/download/v0.6.0/auto_gptq-0.6.0+rocm5.6-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/jllllll/AutoGPTQ/releases/download/v0.6.0/auto_gptq-0.6.0+rocm5.6-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
https://github.com/oobabooga/exllamav2/releases/download/v0.0.20/exllamav2-0.0.20+rocm5.6-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/oobabooga/exllamav2/releases/download/v0.0.20/exllamav2-0.0.20+rocm5.6-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
https://github.com/oobabooga/exllamav2/releases/download/v0.0.20/exllamav2-0.0.20-py3-none-any.whl; platform_system != "Darwin" and platform_machine != "x86_64"
https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+rocm5.6-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+rocm5.6-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
https://github.com/casper-hansen/AutoAWQ/releases/download/v0.2.3/autoawq-0.2.3+rocm561-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/casper-hansen/AutoAWQ/releases/download/v0.2.3/autoawq-0.2.3+rocm561-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
https://github.com/casper-hansen/AutoAWQ/releases/download/v0.2.5/autoawq-0.2.5+rocm561-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/casper-hansen/AutoAWQ/releases/download/v0.2.5/autoawq-0.2.5+rocm561-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"

14
requirements_amd_noavx2.txt
View File

@ -1,9 +1,9 @@
accelerate==0.27.*
accelerate==0.30.*
colorama
datasets
einops
gradio==4.26.*
hqq==0.1.5
hqq==0.1.7.post2
jinja2==3.1.2
lm_eval==0.3.0
markdown
@ -21,7 +21,7 @@ safetensors==0.4.*
scipy
sentencepiece
tensorboard
transformers==4.40.*
transformers==4.41.*
tqdm
wandb
@ -38,12 +38,8 @@ https://github.com/oobabooga/llama-cpp-python-cuBLAS-wheels/releases/download/cp
https://github.com/oobabooga/llama-cpp-python-cuBLAS-wheels/releases/download/cpu/llama_cpp_python-0.2.75+cpuavx-cp310-cp310-win_amd64.whl; platform_system == "Windows" and python_version == "3.10"
# AMD wheels
https://github.com/jllllll/AutoGPTQ/releases/download/v0.6.0/auto_gptq-0.6.0+rocm5.6-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/jllllll/AutoGPTQ/releases/download/v0.6.0/auto_gptq-0.6.0+rocm5.6-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
https://github.com/oobabooga/exllamav2/releases/download/v0.0.20/exllamav2-0.0.20+rocm5.6-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/oobabooga/exllamav2/releases/download/v0.0.20/exllamav2-0.0.20+rocm5.6-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
https://github.com/oobabooga/exllamav2/releases/download/v0.0.20/exllamav2-0.0.20-py3-none-any.whl; platform_system != "Darwin" and platform_machine != "x86_64"
https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+rocm5.6-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+rocm5.6-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
https://github.com/casper-hansen/AutoAWQ/releases/download/v0.2.3/autoawq-0.2.3+rocm561-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/casper-hansen/AutoAWQ/releases/download/v0.2.3/autoawq-0.2.3+rocm561-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
https://github.com/casper-hansen/AutoAWQ/releases/download/v0.2.5/autoawq-0.2.5+rocm561-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/casper-hansen/AutoAWQ/releases/download/v0.2.5/autoawq-0.2.5+rocm561-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"

6
requirements_apple_intel.txt
View File

@ -1,9 +1,9 @@
accelerate==0.27.*
accelerate==0.30.*
colorama
datasets
einops
gradio==4.26.*
hqq==0.1.5
hqq==0.1.7.post2
jinja2==3.1.2
lm_eval==0.3.0
markdown
@ -21,7 +21,7 @@ safetensors==0.4.*
scipy
sentencepiece
tensorboard
transformers==4.40.*
transformers==4.41.*
tqdm
wandb

6
requirements_apple_silicon.txt
View File

@ -1,9 +1,9 @@
accelerate==0.27.*
accelerate==0.30.*
colorama
datasets
einops
gradio==4.26.*
hqq==0.1.5
hqq==0.1.7.post2
jinja2==3.1.2
lm_eval==0.3.0
markdown
@ -21,7 +21,7 @@ safetensors==0.4.*
scipy
sentencepiece
tensorboard
transformers==4.40.*
transformers==4.41.*
tqdm
wandb

6
requirements_cpu_only.txt
View File

@ -1,9 +1,9 @@
accelerate==0.27.*
accelerate==0.30.*
colorama
datasets
einops
gradio==4.26.*
hqq==0.1.5
hqq==0.1.7.post2
jinja2==3.1.2
lm_eval==0.3.0
markdown
@ -21,7 +21,7 @@ safetensors==0.4.*
scipy
sentencepiece
tensorboard
transformers==4.40.*
transformers==4.41.*
tqdm
wandb

6
requirements_cpu_only_noavx2.txt
View File

@ -1,9 +1,9 @@
accelerate==0.27.*
accelerate==0.30.*
colorama
datasets
einops
gradio==4.26.*
hqq==0.1.5
hqq==0.1.7.post2
jinja2==3.1.2
lm_eval==0.3.0
markdown
@ -21,7 +21,7 @@ safetensors==0.4.*
scipy
sentencepiece
tensorboard
transformers==4.40.*
transformers==4.41.*
tqdm
wandb

19
requirements_noavx2.txt
View File

@ -1,11 +1,12 @@
accelerate==0.27.*
aqlm[gpu,cpu]==1.1.3; platform_system == "Linux"
accelerate==0.30.*
aqlm[gpu,cpu]==1.1.5; platform_system == "Linux"
auto-gptq==0.7.1
bitsandbytes==0.43.*
colorama
datasets
einops
gradio==4.26.*
hqq==0.1.5
hqq==0.1.7.post2
jinja2==3.1.2
lm_eval==0.3.0
markdown
@ -23,7 +24,7 @@ safetensors==0.4.*
scipy
sentencepiece
tensorboard
transformers==4.40.*
transformers==4.41.*
tqdm
wandb
@ -52,10 +53,6 @@ https://github.com/oobabooga/llama-cpp-python-cuBLAS-wheels/releases/download/te
https://github.com/oobabooga/llama-cpp-python-cuBLAS-wheels/releases/download/textgen-webui/llama_cpp_python_cuda_tensorcores-0.2.75+cu121avx-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
# CUDA wheels
https://github.com/jllllll/AutoGPTQ/releases/download/v0.6.0/auto_gptq-0.6.0+cu121-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
https://github.com/jllllll/AutoGPTQ/releases/download/v0.6.0/auto_gptq-0.6.0+cu121-cp310-cp310-win_amd64.whl; platform_system == "Windows" and python_version == "3.10"
https://github.com/jllllll/AutoGPTQ/releases/download/v0.6.0/auto_gptq-0.6.0+cu121-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/jllllll/AutoGPTQ/releases/download/v0.6.0/auto_gptq-0.6.0+cu121-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
https://github.com/oobabooga/exllamav2/releases/download/v0.0.20/exllamav2-0.0.20+cu121-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
https://github.com/oobabooga/exllamav2/releases/download/v0.0.20/exllamav2-0.0.20+cu121-cp310-cp310-win_amd64.whl; platform_system == "Windows" and python_version == "3.10"
https://github.com/oobabooga/exllamav2/releases/download/v0.0.20/exllamav2-0.0.20+cu121-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
@ -65,8 +62,4 @@ https://github.com/oobabooga/flash-attention/releases/download/v2.5.6/flash_attn
https://github.com/oobabooga/flash-attention/releases/download/v2.5.6/flash_attn-2.5.6+cu122torch2.2.0cxx11abiFALSE-cp310-cp310-win_amd64.whl; platform_system == "Windows" and python_version == "3.10"
https://github.com/Dao-AILab/flash-attention/releases/download/v2.5.6/flash_attn-2.5.6+cu122torch2.2cxx11abiFALSE-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/Dao-AILab/flash-attention/releases/download/v2.5.6/flash_attn-2.5.6+cu122torch2.2cxx11abiFALSE-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+cu121-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+cu121-cp310-cp310-win_amd64.whl; platform_system == "Windows" and python_version == "3.10"
https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+cu121-cp311-cp311-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.11"
https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+cu121-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.10"
autoawq==0.2.3; platform_system == "Linux" or platform_system == "Windows"
autoawq==0.2.5; platform_system == "Linux" or platform_system == "Windows"

6
requirements_nowheels.txt
View File

@ -1,9 +1,9 @@
accelerate==0.27.*
accelerate==0.30.*
colorama
datasets
einops
gradio==4.26.*
hqq==0.1.5
hqq==0.1.7.post2
jinja2==3.1.2
lm_eval==0.3.0
markdown
@ -21,7 +21,7 @@ safetensors==0.4.*
scipy
sentencepiece
tensorboard
transformers==4.40.*
transformers==4.41.*
tqdm
wandb

11
server.py
View File

@ -32,7 +32,7 @@ import sys
import time
from functools import partial
from pathlib import Path
from threading import Lock
from threading import RLock, Thread
import yaml
@ -52,7 +52,7 @@ from modules import (
)
from modules.extensions import apply_extensions
from modules.LoRA import add_lora_to_model
from modules.models import load_model
from modules.models import load_model, unload_model_if_idle
from modules.models_settings import (
get_fallback_settings,
get_model_metadata,
@ -243,7 +243,12 @@ if __name__ == "__main__":
if shared.args.lora:
add_lora_to_model(shared.args.lora)
shared.generation_lock = Lock()
shared.generation_lock = RLock()
if shared.args.idle_timeout > 0:
timer_thread = Thread(target=unload_model_if_idle)
timer_thread.daemon = True
timer_thread.start()
if shared.args.nowebui:
# Start the API in standalone mode